wehub-resource-sync
124 lines
4.7 KiB
TypeScript
124 lines
4.7 KiB
TypeScript
'use client';
|
|
|
|
/**
|
|
* Single source of truth for "an agent's TTS voice".
|
|
*
|
|
* Every TTS path (lecture narration, multi-agent discussion, voice preview,
|
|
* settings test) resolves voice options through `resolveAgentVoiceOptions`,
|
|
* which reads the agent profile (`voiceConfig` + `voiceDesign`) and, for
|
|
* registration-capable providers, ensures the auto voice is registered and
|
|
* referenced by id (stable timbre) — otherwise falls back to the inline
|
|
* voice-design prompt. There is no second code path to drift out of sync.
|
|
*/
|
|
|
|
import type { AgentConfig } from '@/lib/orchestration/registry/types';
|
|
import type { VoiceDesign } from '@/lib/audio/voice-design';
|
|
import { getVoxCPMProviderOptions } from '@/lib/audio/voxcpm-voices';
|
|
import {
|
|
VOXCPM_AUTO_VOICE_ID,
|
|
VOXCPM_TTS_PROVIDER_ID,
|
|
normalizeVoxCPMBackend,
|
|
} from '@/lib/audio/voxcpm';
|
|
import { useSettingsStore } from '@/lib/store/settings';
|
|
|
|
interface TTSProviderConfigShape {
|
|
apiKey?: string;
|
|
baseUrl?: string;
|
|
customDefaultBaseUrl?: string;
|
|
modelId?: string;
|
|
providerOptions?: Record<string, unknown>;
|
|
}
|
|
|
|
export interface AgentVoiceResolveOptions {
|
|
providerId: string;
|
|
providerConfig?: TTSProviderConfigShape;
|
|
voiceId: string;
|
|
/** Course language / locale — only selects the one-time bootstrap sample sentence. */
|
|
language?: string;
|
|
}
|
|
|
|
/**
|
|
* Pick the agent whose voice narration should use (the teacher).
|
|
*
|
|
* The registry is always seeded with the DEFAULT agents, so a plain
|
|
* `find(role === 'teacher')` returns the default teacher (no voiceDesign) even
|
|
* when a generated classroom is active. Prefer a teacher that actually carries a
|
|
* voiceDesign (the generated one) so narration registers a real voice instead of
|
|
* drifting on the inline prompt; fall back to any teacher.
|
|
*/
|
|
export function pickNarratorAgent(agents: AgentConfig[]): AgentConfig | undefined {
|
|
return (
|
|
agents.find((a) => a.role === 'teacher' && a.voiceDesign) ??
|
|
agents.find((a) => a.role === 'teacher')
|
|
);
|
|
}
|
|
|
|
/**
|
|
* The descriptor used to bootstrap an agent's voice: the real `voiceDesign`
|
|
* when present (generated agents), otherwise the persona as a fallback seed.
|
|
* Persona is not a vocal description, so the resulting voice is stable but
|
|
* generic — good enough to register a consistent reference clip (no drift),
|
|
* pending a real descriptor if quality matters for that agent.
|
|
*/
|
|
function effectiveVoiceDesign(agent: AgentConfig | undefined): VoiceDesign | undefined {
|
|
if (agent?.voiceDesign) return agent.voiceDesign;
|
|
const persona = agent?.persona?.trim();
|
|
return persona ? { identity: persona, texture: '', delivery: '' } : undefined;
|
|
}
|
|
|
|
/**
|
|
* Produce the `ttsProviderOptions` to send to /api/generate/tts for `agent`
|
|
* (pass the teacher agent for narration, the speaking agent for discussion,
|
|
* or undefined when there is no agent). Returns undefined for providers with
|
|
* no special options.
|
|
*/
|
|
export async function resolveAgentVoiceOptions(
|
|
agent: AgentConfig | undefined,
|
|
opts: AgentVoiceResolveOptions,
|
|
): Promise<Record<string, unknown> | undefined> {
|
|
if (opts.providerId !== VOXCPM_TTS_PROVIDER_ID) return undefined;
|
|
return {
|
|
...(opts.providerConfig?.providerOptions || {}),
|
|
...(await getVoxCPMProviderOptions(
|
|
opts.voiceId,
|
|
{
|
|
agentName: agent?.name,
|
|
role: agent?.role ?? 'teacher',
|
|
persona: agent?.persona,
|
|
voiceDesign: effectiveVoiceDesign(agent),
|
|
language: opts.language,
|
|
backend: normalizeVoxCPMBackend(opts.providerConfig?.providerOptions?.backend),
|
|
},
|
|
{
|
|
ttsApiKey: opts.providerConfig?.apiKey || undefined,
|
|
ttsBaseUrl:
|
|
opts.providerConfig?.baseUrl || opts.providerConfig?.customDefaultBaseUrl || undefined,
|
|
ttsModelId: opts.providerConfig?.modelId,
|
|
},
|
|
)),
|
|
};
|
|
}
|
|
|
|
/**
|
|
* Eager warm-up: right after generated agents are saved, pre-register the
|
|
* narrator's (teacher's) auto voice using the SAME idempotent ensure as the TTS
|
|
* path, so the first spoken line is already stable. Only the narrator is warmed:
|
|
* it always speaks (lecture narration), whereas discussion agents may never be
|
|
* selected, so warming all of them would synthesize voices that go unused.
|
|
* Fire-and-forget; the on-use ensure remains the correctness path for the rest.
|
|
*/
|
|
export function warmUpAgentVoices(agents: AgentConfig[]): void {
|
|
const settings = useSettingsStore.getState();
|
|
const providerId = settings.ttsProviderId;
|
|
if (providerId !== VOXCPM_TTS_PROVIDER_ID) return;
|
|
const providerConfig = settings.ttsProvidersConfig?.[providerId];
|
|
|
|
const narrator = pickNarratorAgent(agents);
|
|
if (!narrator || !effectiveVoiceDesign(narrator)) return;
|
|
void resolveAgentVoiceOptions(narrator, {
|
|
providerId,
|
|
providerConfig,
|
|
voiceId: VOXCPM_AUTO_VOICE_ID,
|
|
}).catch(() => undefined);
|
|
}
|